AI review rounds
Human in the loop AI review
Human in the loop AI review means a person signs off on what the models produce before it counts. In Evriq, a review round has several models check the same work independently, and then the person named as reviewer approves it or sends it back. The models recommend and people decide.
How a review round works
Start a round from the session header with Review, or type /review in the session. Every seat in the session takes part, whether it runs on Claude, Codex, Gemini, DeepSeek or Grok.
The round plays out in the transcript where everyone on the project can read it. Its verdict is recorded on the session and written to the audit log. If someone sends a new message while a round is open, the round is cancelled so it never answers a stale question.
Blind answers, then revision
First, every seat answers the same question blind. The blind prompt carries the question and what came before it, and no other seat's answer, so each model works on its own.
Then each seat sees the others' answers once and may revise its own. That single revision pass lets a model correct itself after reading a point it missed, without the round turning into a long back and forth.
The verdict: approve, request changes, reject
One seat writes the verdict, with its reasons and what must change. The session's lead chooses which seat writes it. If a seat leaves out its verdict line, the next seat is asked for one and the transcript says why.
The verdict says how many independent seats answered, counted one per provider. A round where only one model took part is labelled a single-model check when it opens and in its verdict, and the review dialog offers a seat on another provider.
A person approves the work item
A work item is one piece of work in a project, such as a matter, a deal, a ticket, a campaign or a repo change. It has a person responsible and a named reviewer, and it moves from open to in review to approved to done.
The person responsible sends the item for review. Its reviewer, or an administrator, approves it or sends it back with reasons, and the person responsible is told either way. The models' verdict is shown as a recommendation and never worded as the approval. Read more in review and approval.
Decisions agents leave for people
When an agent reaches something only a person can decide, it files a decision with its recommendation, often addressed to one named person. The decision lands in the team's shared inbox, and the answer goes back to the session that asked.
Your inbox badge counts what is asked of you. A decision past its due day is marked overdue, and you can extend it a week or close it. See work items and decisions.
When a single model is not enough
One model reviewing its own work tends to agree with itself. Models from different providers were trained differently and reason differently, so a blind answer from each puts independent checks in front of the person who signs off.
Use a round for a contract clause, a pricing change, a support reply that sets policy, or a code change before it ships. Running several providers in one session is covered on multi-model AI chat. Review rounds count against your workspace's turn and spend limits like any other turn.
Questions about AI review
What is human in the loop AI?
It is AI work where a person reviews and approves the result before it counts. In Evriq, models can review each other in a round, and the work item's named reviewer makes the approval.
Can several AI models check each other's answers?
Yes. In a review round every seat answers blind, then sees the others and may revise once, and one seat writes a verdict with reasons.
Does a review round remove hallucinations?
It puts independent checks in front of a person. Each model answers without seeing the others, so a claim one model gets wrong has a better chance of being caught before the reviewer signs off.
Can I try a review round for free?
Yes. The trial is 14 days with no card needed to start. Start a free trial or open the demo.